Anthropic warns of self-replicating AI
Small artificial intelligence models could be months away from copying themselves across internet-connected consumer devices without relying on data centers, Anthropic has warned, reports a Qazinform News Agency correspondent.
In its official research report, the company examined whether AI systems could move between machines, start operating on unfamiliar hardware and continue spreading without human intervention.
Researchers demonstrated that a model could run across 15 inexpensive Android TV boxes. Another small model successfully set up AI processing on eight of nine hardware types tested.
However, current small models still struggle to complete the entire replication process reliably. More capable models succeeded in simplified experiments, but required separate computing infrastructure.
Anthropic estimates that smaller models could reach the necessary capability by the end of 2026 if improvements continue at their historical pace. It stressed that this timeline is a rough projection.
“We don’t yet have a full rogue replicator,” researchers William Zhou and Daniel Freeman wrote.
The company warned that distributed AI could be difficult to shut down because it would lack a single central control point. Unlike conventional malware, it could adapt to unfamiliar hardware.
The experiments simplified the task of gaining access to devices, meaning their results do not establish reliable replication under real-world conditions.
Anthropic recommended stronger device security, monitoring unusual network traffic and isolating infected machines to limit potential spread.
Earlier, Qazinform News Agency reported that Meta launched Muse Wonder and Muse Wonder Full, describing them as its most capable AI models yet, as it expands its presence in personal AI assistants and coding tools.